Routing, resolution, failures and cost

Run the 60 conversations with tracing on. Score routing accuracy, resolution rate, correct escalations, turns, cost and time. Read traces of failures and label them: wrong route, endless hand-off loop, agent ignored shared state, tool error, made-up facts. Fix and re-run. Then cut cost: a smaller model for routing, caching, limits on hand-offs and steps — and show the before/after table.